Brain Region Segmentation — Human Brain (Allen Depth-3)
DINOv2-Large + UperNet model fine-tuned for semantic segmentation of
human brain regions in Nissl-stained histological sections.
Model Details
| Attribute | Value |
|---|
| Architecture | DINOv2-Large (304M) + UperNet (38M) |
| Classes | 44 (depth-3 brain regions) |
| Input Size | 518x518 |
| Training Data | Allen Human Brain Atlas (6 donors, Nissl staining) |
| mIoU (val center-crop) | 65.5% |
| mIoU (test sliding window) | 65.0% |
| Pixel accuracy (test) | 99.1% |
Usage
1git clone https://github.com/Noel-Niko/histological-image-analysis
2cd histological-image-analysis
3make install
4make download-models-human-allen
5make annotate-human-allen IMAGES=/path/to/your/slides/
Paper
Cross-Species Transfer of Ultra-Fine-Grained Brain Segmentation: From Mouse to Human with DINOv2 + UperNet
We extend the DINOv2-Large + UperNet approach from mouse (1,328 classes, 79.1% mIoU) to human brain tissue using the Allen Human Brain Atlas (sparse SVG annotations, 6 donors). The depth-3 model (44 brain regions) achieves 65.5% val CC mIoU and 65.0% test SW mIoU with 99.1% pixel accuracy. Major structures (cerebellum, cerebral cortex, thalamus, pons) exceed 99% IoU.
See
paper.md in this repo for the full paper.
Citation
If you use this model, please cite the training data sources and the paper included in this repository.
Repository
Maintaining This Repo
To update model weights, papers, or this README:
1cd histological-image-analysis
2export HUGGING_FACE_TOKEN=hf_your_token_here
3
4# Update model weights (Databricks or local):
5jupyter notebook notebooks/upload_models_to_hf.ipynb
6
7# Update papers + READMEs (local only):
8jupyter notebook notebooks/upload_papers_to_hf.ipynb
